I can see a day where the PID loops and basic control is done at the controller level, and the optimizations or the basic algorithms or the advanced algorithms are running at the edge on a bit of a, you know, better, PC, and then heavy machine learning, model building, advanced fault detection is done in the cloud. We think that's probably the progression.
âTerry Herr on the Nexus Podcast, Episode #009
Good morning!
Welcome to Nexus, a newsletter, podcast, and membership community for smart people applying smart building technologyâwritten by James Dice. If youâre new to Nexus, you might want to start here.
Hereâs an outline of this weekâs newsletter:
đ¤ ON MY MIND
đ¤ MACHINE LEARNING FOR BUILDINGS 101
đĄ NEW FROM NEXUS
đ WHAT IâM READING
Enjoy!
Oh, and by the way: if you missed last weekâs edition, you can find it here.
Disclaimer: James is a researcher at the National Renewable Energy Laboratory (NREL). All opinions expressed via Nexus emails, podcasts, or on the website belong solely to James. No resources from NREL are used to support Nexus. NREL does not endorse or support any aspect of Nexus.
1. đ¤ ON MY MIND
A reader asked me last week whether I think AI (for buildings) is hype or not. Thatâs a very difficult question to answer!
First, our industry has a lot of hype. Thatâs part of the reason Nexus exists. So my answer to these types of questions is usually a resounding âYESâ.
Second, is anyone on the same page with the term âAIâ? âMachine learningâ? I donât think so. My favorite is when someone says they have âAI and MLâ. My second favorite is when someone talks about it like a chef adding salt.
As always, cutting through hype means defining the use case. So thatâs where weâre going to start.
Iâd love your feedback on where we can take this and what misconceptions need to be cleared up.
2.đ¤ MACHINE LEARNING FOR BUILDINGS 101
In our industry, as in the broader media, we see terms like machine learning, artificial intelligence, deep learning, etc used interchangeably. Are they all the same thing? We also see them used as discrete, parallel advancements to create hype around new products and companies. Are they all different?
No and no. Itâs better to think of them as concentric circles. Or, in the case of this graphic from NVIDIA, concentric rectangles:
Deep learning is a type of machine learning which is a type of artificial intelligence. Some general definitions:
AI is a broader term for human intelligence exhibited by machines
ML is an approach to achieve AI by teaching a computer to recognize patterns in data by example and make predictions
DL is a type of machine learning that uses neural networks with many layers
The breakthrough created by machine learning is in the scalability and portability of algorithms. Rather than hand-coding software routines with a specific set of instructions to accomplish a particular task, the machine is âtrainedâ using large amounts of data and algorithms that give it the ability to learn how to perform the task on its own.
For example, consider Google Photos. If I open the app and search âcatâ, it will return all of my pictures of my cat Grace. Before machine learning, the programmers at Google would have needed to write special code for recognizing cats. But if I wanted to search âdogâ or âMomâ or âcampingâ or ârock climbingâ, that special code would not have been useful. Thanks to machine learning, one system can learn to recognize anything if we show it enough examples of that thing.
In general, most applications for AI in buildings are actually machine learning and most ML advancements in recent years are in DL. The exception to this generalization, as my friend Alex Grace told me on Episode 4 of the Nexus podcast, is hierarchical, rule-based fault detection and diagnostics (FDD). This is a type of AI called an expert system. It is AI because it can mimic human intelligence, but is not a type of ML because it doesnât learn to do so by itself.
USE CASES FOR MACHINE LEARNING IN BUILDINGS
That brings us to use cases. While Iâve seen a lot of great resources, I havenât yet seen a concise, introductory list of the use cases for machine learning in buildings. I took a crack at starting that list below.
To move the industry forward, machine learning needs to either enable us to perform new tasks or perform old tasks better. Understanding the three different types of machine learning (supervised, unsupervised, and reinforcement) helps us understand how it actually works and compare it to old ways of doing things.
Letâs walk through each type and use cases for each:
Supervised learning is when the system is taught using training data that is labeled. The system is presented with example inputs and their desired outputs and the goal is to learn a model that maps inputs to outputs. Use cases:
Forecasting peak demand (or other variables) based on weather forecasts and the predicted building system response
Normalizing for energy savings measurement & verification to account for changes in independent variables (weather, occupancy, etc)
Disaggregation/itemization of whole-building energy meter data into end uses
Automated tagging and semantic modeling of data points (typically combined with rule-based expert systems and unsupervised learning)
Predictive maintenance (see Augury; also could use unsupervised learning if no failure data is available)
Counting people with sensors or cameras (See Density)
Interpreting camera feeds (e.g. to check for PPE or social distancing compliance)
Speech recognition (think Amazon Alexa for buildings)
Unsupervised learning is when the system is taught using unlabeled data. It is on its own to find structure in its input. The goal may be to discover hidden patterns in data or to create groups according to similar features. Use cases:
Setting (or recommending) dynamic equipment and building schedules based on historical occupancy patterns
Automatically detecting anomalies in any time series data point (another way to do FDD; see LeanFM)
Detecting correlation and inferring relationships between equipment (see Nexus #2)
Learning the energy flows of each zone in the building and then predicting their future state (see Nexus #18)
Reinforcement learning is when the system automatically performs a certain task (such as driving a vehicle or playing a game against an opponent) and is provided feedback in terms of rewards and punishments. Use cases:
Controlling a building while optimizing across many variables and constraints, such as when to charge a battery or when to precool to minimize demand (see this deep dive).
Obviously this is just a start, but I plan to keep the list updated as I come across new use cases. What would you add to the list?
3. đĄ NEW FROM NEXUS
PODCASTâEpisode #009 of the Nexus Podcast is a conversation with Terry Herr, President of Intellimation, on the past, present, and future of analytics for buildings.
DEEP DIVEâAfter the interview with Terry, I did a deep dive on my reaction, my top highlights of the episode, and a full transcript (Pro members only)
EVENTâJuneâs member gathering is on the calendar! Pro members already received a calendar invite. Hereâs the plan:
Weâll do two breakout rooms so you can meet likeminded industry leaders
Dennis Krieger, Director of Engineering at Willow, will present on âDemystifying the Digital Twin: Connecting BIM with IoTâ
âŚonce I had started thinking about the legacy the real estate industry had on racial injustice, I couldnât stop researching it. I kept finding more and more shocking examples of institutional racism, ones that seemed to have slipped out of the positive narrative that the real estate industry likes to focus on.
The biggest change will be the separation of technologies in buildings. There is an expectation that building automation systems and the equipment that people use in a facility are separate. Typically, in a boardroom, an access system controls entrance into the space if access is automated. It is booked for use through proprietary software at a PC. Temperature is controlled by a wall thermostat.
Users adjust lighting by common switches and dimmers. Automatic blinds are controlled by remotes or touchpads. AV equipment is manually switched on by another remote control. All of these are potential infection points when used by multiple people every day. As long as these systems remain separate, they will remain infection points. All of these systems will need to become more commonly integrated to have a meaningful effect on the safety of office spaces.
Before we review the relevant studies and draw out lessons for the future of the great indoors, a brief word of humility. Our understanding of this disease is dynamic. Todayâs conventional wisdom could be tomorrowâs busted myth. Think of these studies not as gospels, but as clues in a gradually unraveling mystery.
one of the newest features of the technology comes in the form of installing digital screens outside cafeterias, meeting rooms and other spaces, allowing occupiers to see for themselves how close to an unsafe density a given room is.
Actually, I was underwhelmed by this one, despite being in agreement on the utility of occupancy sensors. I still decided to share it because I want to call out this trend (or perhaps pet peeve): this author and others mention âoccupancy sensorsâ as some sort of stand-alone solution when in fact it is not. Not even close. Do you agree?
businesses that are typically reluctant to spend on replacing older technology have indicated that significant funding is available for contact tracing in the workplace.
If youâre an engineer and you read this newsletter and youâre looking for a new job⌠Slipstream is hiring:
OK, thatâs all for this weekâthanks for reading Nexus!
âJames
P.S. Nexus is a 100 percent reader-funded publication and podcast. It only exists because people like you support it. If you liked todayâs edition, please consider joining Nexus Pro! Members get exclusive access to the Nexus Vendor Landscape, monthly events, weekly deep dives, and all past deep dives like these:
I can see a day where the PID loops and basic control is done at the controller level, and the optimizations or the basic algorithms or the advanced algorithms are running at the edge on a bit of a, you know, better, PC, and then heavy machine learning, model building, advanced fault detection is done in the cloud. We think that's probably the progression.
âTerry Herr on the Nexus Podcast, Episode #009
Good morning!
Welcome to Nexus, a newsletter, podcast, and membership community for smart people applying smart building technologyâwritten by James Dice. If youâre new to Nexus, you might want to start here.
Hereâs an outline of this weekâs newsletter:
đ¤ ON MY MIND
đ¤ MACHINE LEARNING FOR BUILDINGS 101
đĄ NEW FROM NEXUS
đ WHAT IâM READING
Enjoy!
Oh, and by the way: if you missed last weekâs edition, you can find it here.
Disclaimer: James is a researcher at the National Renewable Energy Laboratory (NREL). All opinions expressed via Nexus emails, podcasts, or on the website belong solely to James. No resources from NREL are used to support Nexus. NREL does not endorse or support any aspect of Nexus.
1. đ¤ ON MY MIND
A reader asked me last week whether I think AI (for buildings) is hype or not. Thatâs a very difficult question to answer!
First, our industry has a lot of hype. Thatâs part of the reason Nexus exists. So my answer to these types of questions is usually a resounding âYESâ.
Second, is anyone on the same page with the term âAIâ? âMachine learningâ? I donât think so. My favorite is when someone says they have âAI and MLâ. My second favorite is when someone talks about it like a chef adding salt.
As always, cutting through hype means defining the use case. So thatâs where weâre going to start.
Iâd love your feedback on where we can take this and what misconceptions need to be cleared up.
2.đ¤ MACHINE LEARNING FOR BUILDINGS 101
In our industry, as in the broader media, we see terms like machine learning, artificial intelligence, deep learning, etc used interchangeably. Are they all the same thing? We also see them used as discrete, parallel advancements to create hype around new products and companies. Are they all different?
No and no. Itâs better to think of them as concentric circles. Or, in the case of this graphic from NVIDIA, concentric rectangles:
Deep learning is a type of machine learning which is a type of artificial intelligence. Some general definitions:
AI is a broader term for human intelligence exhibited by machines
ML is an approach to achieve AI by teaching a computer to recognize patterns in data by example and make predictions
DL is a type of machine learning that uses neural networks with many layers
The breakthrough created by machine learning is in the scalability and portability of algorithms. Rather than hand-coding software routines with a specific set of instructions to accomplish a particular task, the machine is âtrainedâ using large amounts of data and algorithms that give it the ability to learn how to perform the task on its own.
For example, consider Google Photos. If I open the app and search âcatâ, it will return all of my pictures of my cat Grace. Before machine learning, the programmers at Google would have needed to write special code for recognizing cats. But if I wanted to search âdogâ or âMomâ or âcampingâ or ârock climbingâ, that special code would not have been useful. Thanks to machine learning, one system can learn to recognize anything if we show it enough examples of that thing.
In general, most applications for AI in buildings are actually machine learning and most ML advancements in recent years are in DL. The exception to this generalization, as my friend Alex Grace told me on Episode 4 of the Nexus podcast, is hierarchical, rule-based fault detection and diagnostics (FDD). This is a type of AI called an expert system. It is AI because it can mimic human intelligence, but is not a type of ML because it doesnât learn to do so by itself.
USE CASES FOR MACHINE LEARNING IN BUILDINGS
That brings us to use cases. While Iâve seen a lot of great resources, I havenât yet seen a concise, introductory list of the use cases for machine learning in buildings. I took a crack at starting that list below.
To move the industry forward, machine learning needs to either enable us to perform new tasks or perform old tasks better. Understanding the three different types of machine learning (supervised, unsupervised, and reinforcement) helps us understand how it actually works and compare it to old ways of doing things.
Letâs walk through each type and use cases for each:
Supervised learning is when the system is taught using training data that is labeled. The system is presented with example inputs and their desired outputs and the goal is to learn a model that maps inputs to outputs. Use cases:
Forecasting peak demand (or other variables) based on weather forecasts and the predicted building system response
Normalizing for energy savings measurement & verification to account for changes in independent variables (weather, occupancy, etc)
Disaggregation/itemization of whole-building energy meter data into end uses
Automated tagging and semantic modeling of data points (typically combined with rule-based expert systems and unsupervised learning)
Predictive maintenance (see Augury; also could use unsupervised learning if no failure data is available)
Counting people with sensors or cameras (See Density)
Interpreting camera feeds (e.g. to check for PPE or social distancing compliance)
Speech recognition (think Amazon Alexa for buildings)
Unsupervised learning is when the system is taught using unlabeled data. It is on its own to find structure in its input. The goal may be to discover hidden patterns in data or to create groups according to similar features. Use cases:
Setting (or recommending) dynamic equipment and building schedules based on historical occupancy patterns
Automatically detecting anomalies in any time series data point (another way to do FDD; see LeanFM)
Detecting correlation and inferring relationships between equipment (see Nexus #2)
Learning the energy flows of each zone in the building and then predicting their future state (see Nexus #18)
Reinforcement learning is when the system automatically performs a certain task (such as driving a vehicle or playing a game against an opponent) and is provided feedback in terms of rewards and punishments. Use cases:
Controlling a building while optimizing across many variables and constraints, such as when to charge a battery or when to precool to minimize demand (see this deep dive).
Obviously this is just a start, but I plan to keep the list updated as I come across new use cases. What would you add to the list?
3. đĄ NEW FROM NEXUS
PODCASTâEpisode #009 of the Nexus Podcast is a conversation with Terry Herr, President of Intellimation, on the past, present, and future of analytics for buildings.
DEEP DIVEâAfter the interview with Terry, I did a deep dive on my reaction, my top highlights of the episode, and a full transcript (Pro members only)
EVENTâJuneâs member gathering is on the calendar! Pro members already received a calendar invite. Hereâs the plan:
Weâll do two breakout rooms so you can meet likeminded industry leaders
Dennis Krieger, Director of Engineering at Willow, will present on âDemystifying the Digital Twin: Connecting BIM with IoTâ
âŚonce I had started thinking about the legacy the real estate industry had on racial injustice, I couldnât stop researching it. I kept finding more and more shocking examples of institutional racism, ones that seemed to have slipped out of the positive narrative that the real estate industry likes to focus on.
The biggest change will be the separation of technologies in buildings. There is an expectation that building automation systems and the equipment that people use in a facility are separate. Typically, in a boardroom, an access system controls entrance into the space if access is automated. It is booked for use through proprietary software at a PC. Temperature is controlled by a wall thermostat.
Users adjust lighting by common switches and dimmers. Automatic blinds are controlled by remotes or touchpads. AV equipment is manually switched on by another remote control. All of these are potential infection points when used by multiple people every day. As long as these systems remain separate, they will remain infection points. All of these systems will need to become more commonly integrated to have a meaningful effect on the safety of office spaces.
Before we review the relevant studies and draw out lessons for the future of the great indoors, a brief word of humility. Our understanding of this disease is dynamic. Todayâs conventional wisdom could be tomorrowâs busted myth. Think of these studies not as gospels, but as clues in a gradually unraveling mystery.
one of the newest features of the technology comes in the form of installing digital screens outside cafeterias, meeting rooms and other spaces, allowing occupiers to see for themselves how close to an unsafe density a given room is.
Actually, I was underwhelmed by this one, despite being in agreement on the utility of occupancy sensors. I still decided to share it because I want to call out this trend (or perhaps pet peeve): this author and others mention âoccupancy sensorsâ as some sort of stand-alone solution when in fact it is not. Not even close. Do you agree?
businesses that are typically reluctant to spend on replacing older technology have indicated that significant funding is available for contact tracing in the workplace.
If youâre an engineer and you read this newsletter and youâre looking for a new job⌠Slipstream is hiring:
OK, thatâs all for this weekâthanks for reading Nexus!
âJames
P.S. Nexus is a 100 percent reader-funded publication and podcast. It only exists because people like you support it. If you liked todayâs edition, please consider joining Nexus Pro! Members get exclusive access to the Nexus Vendor Landscape, monthly events, weekly deep dives, and all past deep dives like these:
Join Nexus Pro and get full access including invite-only member gatherings, access to the community chatroom Nexus Connect, networking opportunities, and deep dive essays.
